Best probe direction is not the best steering direction, Fraunhofer study finds across models
FraunhoferIAIS · hf · 2026-11-20
Fraunhofer IAIS researchers systematically compare probing and steering directions in language models' hidden states. Constructing a one-parameter family of directions from a covariance-adjusted mean difference, they evaluate probing and additive steering at a fixed side effect. For truthfulness and sentiment, in every tested instruction-tuned model, the best probe is not the best steering direction — steering consistently favors directions closer to the mean difference, and the separation holds across 59 of 60 cases, with implications for transfer and abstention.
More from Research
- AutoRAGBench: LLM agents build RAG systems beating the best published human designs — omarsar0 · 2026-11-22
- Hy team pushes back on ultra-sparse MoE: sparsity recovers efficiency, not quality — douglovefishing12 · 2026-11-22
- Gary Marcus: LLMs work better because they borrowed tree search from symbolic AI — GaryMarcus · 2026-11-22
- Caltech's ASTRA robot autonomously builds 1.4m antenna in orbit, a space first — mattbeane · 2026-11-22
- AAAI 2027 NeuroAI 2.0 workshop extends paper deadline to Nov 23, 2026 — naturecomputes · 2026-11-22
- Gary Marcus: universal approximation doesn't mean networks can learn — GaryMarcus · 2026-11-22